Spaces:

ericjedha
/

crazycat

Running

Update app.py

0642309 verified 10 months ago

14.2 kB

import torch
import torch.nn.functional as F
import gradio as gr
import librosa
import numpy as np
import cv2
import timm
import os
import time
import spaces
import plotly.express as px
from huggingface_hub import hf_hub_download
from transformers import (
    AutoProcessor,
    AutoModelForImageTextToText,
    ASTFeatureExtractor,
    ASTForAudioClassification,
    AutoModelForCausalLM,
    AutoTokenizer
)
from moviepy import VideoFileClip

# --- Configuration ---
CATEGORIES = ['affection', 'angry', 'back_off', 'defensive', 'feed_me', 'happy', 'hunt', 'in_heat', 'mother_call', 'pain', 'wants_attention']
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# ==========================================
# 1. CHARGEMENT DES MODÈLES
# ==========================================
def load_models():
    print("📥 Initialisation CatSense v12.13 (Vision Pure Mode)...")
   
    # Modèle VLM
    vlm_id = "HuggingFaceTB/SmolVLM2-256M-Video-Instruct"
    vlm_model = AutoModelForImageTextToText.from_pretrained(
        vlm_id, torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32
    ).to(DEVICE).eval()

    # LLM Juge
    llm_id = "HuggingFaceTB/SmolLM2-135M-Instruct"
    llm_tok = AutoTokenizer.from_pretrained(llm_id)
    llm_model = AutoModelForCausalLM.from_pretrained(
        llm_id, torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32
    ).to(DEVICE).eval()

    # Audio models
    audio_models = {}
    for p, repo, f in [('A', 'ericjedha/pilier_a', 'best_pillar_a_e29_f1_0_9005.pth'),
                       ('B', 'ericjedha/pilier_b', 'best_pillar_b_f1_09103.pth')]:
        path = hf_hub_download(repo_id=repo, filename=f)
        m = timm.create_model("vit_small_patch16_224", num_classes=len(CATEGORIES), in_chans=3)
        m.load_state_dict(torch.load(path, map_location=DEVICE)['model_state_dict'])
        audio_models[p] = m.to(DEVICE).eval()
       
    path_c = hf_hub_download(repo_id="ericjedha/pilier_c", filename="best_pillar_c_ast_v95_2_f1_0_9109.pth")
    model_c = ASTForAudioClassification.from_pretrained("MIT/ast-finetuned-audioset-10-10-0.4593", num_labels=len(CATEGORIES), ignore_mismatched_sizes=True)
    sd = torch.load(path_c, map_location=DEVICE)['model_state_dict']
    model_c.load_state_dict({k.replace('ast.', ''): v for k, v in sd.items()}, strict=False)
    audio_models['C'] = model_c.to(DEVICE).eval()
    audio_models['ast_ext'] = ASTFeatureExtractor.from_pretrained("MIT/ast-finetuned-audioset-10-10-0.4593")
   
    return vlm_model, llm_tok, llm_model, audio_models

# Chargement global
vlm_model, llm_tok, llm_model, audio_models = load_models()

# ==========================================
# 2. JUGE HYBRIDE (règles + LLM)
# ==========================================
def call_peace_judge(audio_ctx, vlm_desc):
    """
    Deterministic + LLM hybrid judge.
    AUDIO dominates when confidence > 30%.
    Vision can refine but never neutralize strong audio signals.
    """

    vlm_lower = vlm_desc.lower()
    audio_upper = audio_ctx.upper()

    # =====================================================
    # 1. HARD AUDIO GUARDRAILS (ABSOLUTE PRIORITY)
    # =====================================================

    if "PAIN" in audio_upper:
        return "The cat is in pain."

    if "ANGRY" in audio_upper:
        return "The cat is angry."

    if "DEFENSIVE" in audio_upper:
        return "The cat is defensive."

    if "BACK_OFF" in audio_upper or "BACKING_OFF" in audio_upper:
        return "The cat is backing off."

    # =====================================================
    # 2. HARD VISUAL OVERRIDES (SAFETY FIRST)
    # =====================================================

    # Aggression / threat display
    if any(x in vlm_lower for x in [
        "front paws raised", "paws raised", "swiping",
        "hissing", "mouth open and tense"
    ]):
        return "The cat is angry."

    # Defensive posture
    if any(x in vlm_lower for x in [
        "arched back", "puffed fur", "ears flat",
        "ears back", "sideways stance"
    ]):
        return "The cat is defensive."

    # Pain indicators
    if any(x in vlm_lower for x in [
        "limping", "hunched", "crouched low",
        "guarding", "withdrawn posture"
    ]):
        return "The cat is in pain."

    # =====================================================
    # 3. POSITIVE / LOW-RISK VISUAL STATES
    # =====================================================

    if any(x in vlm_lower for x in [
        "kneading", "rubbing", "head bumping"
    ]):
        return "The cat is affectionate."

    if any(x in vlm_lower for x in [
        "playful", "rolling", "pouncing"
    ]):
        return "The cat is happy."

    if any(x in vlm_lower for x in [
        "stalking", "tail twitching", "low crawl"
    ]):
        return "The cat is hunting."

    if any(x in vlm_lower for x in [
        "approaching human", "following human",
        "pawing at leg"
    ]):
        return "The cat is wanting attention."

    if any(x in vlm_lower for x in [
        "waiting posture", "looking at food",
        "pacing near bowl"
    ]):
        return "The cat is hungry."

    # =====================================================
    # 4. LLM FALLBACK (NO CALM ALLOWED)
    # =====================================================

    messages = [
        {
            "role": "system",
            "content": (
                "You are a strict cat behavior decision engine.\n"
                "Rules:\n"
                "1. AUDIO has priority over vision.\n"
                "2. You must choose the most conservative interpretation.\n"
                "3. 'calm' is NOT a valid output.\n"
                "4. If unsure, prefer defensive or in pain.\n\n"
                "Allowed outputs ONLY:\n"
                "affectionate, angry, backing off, defensive, hungry, happy, "
                "hunting, in heat, calling kittens, in pain, wanting attention\n\n"
                "Answer format EXACTLY:\n"
                "The cat is [label]."
            )
        },
        {
            "role": "user",
            "content": (
                f"AUDIO SIGNAL (PRIMARY): {audio_ctx}\n"
                f"VISION OBSERVATIONS (SECONDARY): {vlm_desc}\n\n"
                "FINAL DECISION:"
            )
        }
    ]

    input_text = llm_tok.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    inputs = llm_tok(input_text, return_tensors="pt").to(DEVICE)

    with torch.no_grad():
        outputs = llm_model.generate(
            **inputs,
            max_new_tokens=15,
            do_sample=False,
            temperature=0.0,
            pad_token_id=llm_tok.eos_token_id,
            eos_token_id=llm_tok.eos_token_id
        )

    generated = llm_tok.decode(
        outputs[0][inputs["input_ids"].shape[1]:],
        skip_special_tokens=True
    ).lower()

    for cat in CATEGORIES:
        if cat.replace("_", " ") in generated:
            return f"The cat is {cat.replace('_', ' ')}."

    # =====================================================
    # 5. FINAL FAILSAFE (NEVER CALM)
    # =====================================================
    return "The cat is defensive."


# ==========================================
# 3. PIPELINE ANALYSE COMPLETE (CORRIGÉ)
# ==========================================
@spaces.GPU(duration=120)
def analyze_cat_v12_final(video_path):
    if not video_path:
        return "❌ Aucune vidéo.", None

    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    tmp_audio = f"temp_{os.getpid()}_{int(time.time())}.wav"
    start_total = time.time()

    # --------------------------------------------------
    # Helper: clean VLM repetitions (cheap & mobile-safe)
    # --------------------------------------------------
    def clean_vlm_output(text):
        sentences = text.split(". ")
        cleaned = []
        seen = set()
        for s in sentences:
            key = s.strip().lower()
            if key and key not in seen:
                seen.add(key)
                cleaned.append(s.strip())
        return ". ".join(cleaned)

    try:
        # =========================
        # A. AUDIO
        # =========================
        t_0 = time.time()
        clip = VideoFileClip(video_path)
        audio_probs = np.zeros(len(CATEGORIES))

        if clip.audio:
            clip.audio.write_audiofile(tmp_audio, fps=16000, logger=None)
            w, _ = librosa.load(tmp_audio, sr=16000, duration=5.0)
            if len(w) < 48000:
                w = np.pad(w, (0, 48000 - len(w)))

            mel = librosa.feature.melspectrogram(y=w, sr=16000, n_mels=192)
            mel_db = (librosa.power_to_db(mel, ref=np.max) + 40) / 40
            img = cv2.resize(
                (np.vstack([mel_db, np.zeros((10, mel_db.shape[1]))]) * 255).astype(np.uint8),
                (224, 224)
            )

            img_t = (
                torch.tensor(img)
                .unsqueeze(0)
                .repeat(1, 3, 1, 1)
                .float()
                .to(DEVICE) / 255.0
            )

            with torch.no_grad():
                pa = F.softmax(audio_models['A'](img_t), dim=1)
                pb = F.softmax(audio_models['B'](img_t), dim=1)
                ic = audio_models['ast_ext'](
                    w, sampling_rate=16000, return_tensors="pt"
                ).to(DEVICE)
                pc = F.softmax(audio_models['C'](**ic).logits, dim=1)

                audio_probs = (
                    pa * 0.3468 + pb * 0.2762 + pc * 0.3770
                ).cpu().numpy()[0]

        clip.close()
        t_audio = time.time() - t_0

        # =========================
        # B. VISION (VLM STABILISÉ)
        # =========================
        t_1 = time.time()
        vlm_proc = AutoProcessor.from_pretrained(
            "HuggingFaceTB/SmolVLM2-256M-Video-Instruct"
        )

        vlm_prompt = (
            "You are a feline behavior expert.\n"
            "Describe ONLY observable physical features:\n"
            "- ears position\n"
            "- mouth state (open/closed/tense)\n"
            "- tail position or movement\n"
            "- body posture\n"
            "Use short factual sentences.\n"
            "One observation per sentence.\n"
            "Do NOT interpret mood."
        )

        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "video", "path": video_path},
                    {"type": "text", "text": vlm_prompt}
                ]
            }
        ]

        vlm_inputs = vlm_proc.apply_chat_template(
            messages,
            add_generation_prompt=True,
            tokenize=True,
            return_dict=True,
            return_tensors="pt"
        ).to(DEVICE)

        input_length = vlm_inputs["input_ids"].shape[1]

        with torch.no_grad():
            vlm_out = vlm_model.generate(
                **vlm_inputs,
                max_new_tokens=80,
                do_sample=False,
                temperature=0.0,
                repetition_penalty=1.15,     # 🔑 anti-loop
                no_repeat_ngram_size=5,      # 🔑 anti-phrase répétée
                pad_token_id=vlm_proc.tokenizer.eos_token_id,
                eos_token_id=vlm_proc.tokenizer.eos_token_id
            )

        gen_tokens = vlm_out[0][input_length:]
        vlm_clean = vlm_proc.batch_decode(
            [gen_tokens], skip_special_tokens=True
        )[0]

        vlm_clean = vlm_clean.strip().split("\n")[0]
        if vlm_clean.lower().startswith("assistant:"):
            vlm_clean = vlm_clean.split(":", 1)[-1].strip()

        # nettoyage final anti-répétition
        vlm_clean = clean_vlm_output(vlm_clean)

        if torch.cuda.is_available():
            torch.cuda.empty_cache()

        t_vlm = time.time() - t_1

        # =========================
        # C. JUGE
        # =========================
        t_2 = time.time()
        top_idx = np.argmax(audio_probs)
        audio_ctx = f"{CATEGORIES[top_idx].upper()} ({audio_probs[top_idx]*100:.1f}%)"
        judge_decision = call_peace_judge(audio_ctx, vlm_clean)
        t_llm = time.time() - t_2

        # =========================
        # D. VISUELS
        # =========================
        top5 = np.argsort(audio_probs)[-5:][::-1]
        fig = px.bar(
            x=[audio_probs[i] * 100 for i in top5],
            y=[CATEGORIES[i].upper() for i in top5],
            orientation="h",
            title="Top 5 Scores Audio",
            labels={"x": "Probabilité (%)", "y": "Émotion"},
            color=[audio_probs[i] * 100 for i in top5],
            color_continuous_scale="Viridis"
        )
        fig.update_layout(height=400, showlegend=False)

        # =========================
        # E. RAPPORT FINAL
        # =========================
        t_total = time.time() - start_total
        report = f"""⚖️ VERDICT JUGE : {judge_decision}
------------------------------------------
👁️ VISION : {vlm_clean}
📊 AUDIO : {audio_ctx}
⏱️ TEMPS : Audio {t_audio:.2f}s | Vision {t_vlm:.2f}s | Juge {t_llm:.2f}s | Total {t_total:.2f}s"""

        return report, fig

    except Exception as e:
        return f"❌ Erreur : {str(e)}", None

    finally:
        if os.path.exists(tmp_audio):
            try:
                os.remove(tmp_audio)
            except:
                pass


# --- Interface Gradio ---
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🐱 CatSense v12.13 - Vision Pure Mode")
    gr.Markdown("✅ **SmolVLM2-256M** + **SmolLM2-135M Juge** + Audio Ensemble")
    
    with gr.Row():
        with gr.Column():
            video_input = gr.Video(label="Vidéo du chat")
            btn = gr.Button("🚀 ANALYSER", variant="primary", size="lg")
        with gr.Column():
            report_out = gr.Textbox(label="Résultat complet", lines=12, interactive=False)
            chart_out = gr.Plot(label="Distribution des émotions (Audio)")
    
    btn.click(analyze_cat_v12_final, inputs=video_input, outputs=[report_out, chart_out])

demo.launch()

	import torch
	import torch.nn.functional as F
	import gradio as gr
	import librosa
	import numpy as np
	import cv2
	import timm
	import os
	import time
	import spaces
	import plotly.express as px
	from huggingface_hub import hf_hub_download
	from transformers import (
	AutoProcessor,
	AutoModelForImageTextToText,
	ASTFeatureExtractor,
	ASTForAudioClassification,
	AutoModelForCausalLM,
	AutoTokenizer
	)
	from moviepy import VideoFileClip

	# --- Configuration ---
	CATEGORIES = ['affection', 'angry', 'back_off', 'defensive', 'feed_me', 'happy', 'hunt', 'in_heat', 'mother_call', 'pain', 'wants_attention']
	DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

	# ==========================================
	# 1. CHARGEMENT DES MODÈLES
	# ==========================================
	def load_models():
	print("📥 Initialisation CatSense v12.13 (Vision Pure Mode)...")

	# Modèle VLM
	vlm_id = "HuggingFaceTB/SmolVLM2-256M-Video-Instruct"
	vlm_model = AutoModelForImageTextToText.from_pretrained(
	vlm_id, torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32
	).to(DEVICE).eval()

	# LLM Juge
	llm_id = "HuggingFaceTB/SmolLM2-135M-Instruct"
	llm_tok = AutoTokenizer.from_pretrained(llm_id)
	llm_model = AutoModelForCausalLM.from_pretrained(
	llm_id, torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32
	).to(DEVICE).eval()

	# Audio models
	audio_models = {}
	for p, repo, f in [('A', 'ericjedha/pilier_a', 'best_pillar_a_e29_f1_0_9005.pth'),
	('B', 'ericjedha/pilier_b', 'best_pillar_b_f1_09103.pth')]:
	path = hf_hub_download(repo_id=repo, filename=f)
	m = timm.create_model("vit_small_patch16_224", num_classes=len(CATEGORIES), in_chans=3)
	m.load_state_dict(torch.load(path, map_location=DEVICE)['model_state_dict'])
	audio_models[p] = m.to(DEVICE).eval()

	path_c = hf_hub_download(repo_id="ericjedha/pilier_c", filename="best_pillar_c_ast_v95_2_f1_0_9109.pth")
	model_c = ASTForAudioClassification.from_pretrained("MIT/ast-finetuned-audioset-10-10-0.4593", num_labels=len(CATEGORIES), ignore_mismatched_sizes=True)
	sd = torch.load(path_c, map_location=DEVICE)['model_state_dict']
	model_c.load_state_dict({k.replace('ast.', ''): v for k, v in sd.items()}, strict=False)
	audio_models['C'] = model_c.to(DEVICE).eval()
	audio_models['ast_ext'] = ASTFeatureExtractor.from_pretrained("MIT/ast-finetuned-audioset-10-10-0.4593")

	return vlm_model, llm_tok, llm_model, audio_models

	# Chargement global
	vlm_model, llm_tok, llm_model, audio_models = load_models()

	# ==========================================
	# 2. JUGE HYBRIDE (règles + LLM)
	# ==========================================
	def call_peace_judge(audio_ctx, vlm_desc):
	"""
	Deterministic + LLM hybrid judge.
	AUDIO dominates when confidence > 30%.
	Vision can refine but never neutralize strong audio signals.
	"""

	vlm_lower = vlm_desc.lower()
	audio_upper = audio_ctx.upper()

	# =====================================================
	# 1. HARD AUDIO GUARDRAILS (ABSOLUTE PRIORITY)
	# =====================================================

	if "PAIN" in audio_upper:
	return "The cat is in pain."

	if "ANGRY" in audio_upper:
	return "The cat is angry."

	if "DEFENSIVE" in audio_upper:
	return "The cat is defensive."

	if "BACK_OFF" in audio_upper or "BACKING_OFF" in audio_upper:
	return "The cat is backing off."

	# =====================================================
	# 2. HARD VISUAL OVERRIDES (SAFETY FIRST)
	# =====================================================

	# Aggression / threat display
	if any(x in vlm_lower for x in [
	"front paws raised", "paws raised", "swiping",
	"hissing", "mouth open and tense"
	]):
	return "The cat is angry."

	# Defensive posture
	if any(x in vlm_lower for x in [
	"arched back", "puffed fur", "ears flat",
	"ears back", "sideways stance"
	]):
	return "The cat is defensive."

	# Pain indicators
	if any(x in vlm_lower for x in [
	"limping", "hunched", "crouched low",
	"guarding", "withdrawn posture"
	]):
	return "The cat is in pain."

	# =====================================================
	# 3. POSITIVE / LOW-RISK VISUAL STATES
	# =====================================================

	if any(x in vlm_lower for x in [
	"kneading", "rubbing", "head bumping"
	]):
	return "The cat is affectionate."

	if any(x in vlm_lower for x in [
	"playful", "rolling", "pouncing"
	]):
	return "The cat is happy."

	if any(x in vlm_lower for x in [
	"stalking", "tail twitching", "low crawl"
	]):
	return "The cat is hunting."

	if any(x in vlm_lower for x in [
	"approaching human", "following human",
	"pawing at leg"
	]):
	return "The cat is wanting attention."

	if any(x in vlm_lower for x in [
	"waiting posture", "looking at food",
	"pacing near bowl"
	]):
	return "The cat is hungry."

	# =====================================================
	# 4. LLM FALLBACK (NO CALM ALLOWED)
	# =====================================================

	messages = [
	{
	"role": "system",
	"content": (
	"You are a strict cat behavior decision engine.\n"
	"Rules:\n"
	"1. AUDIO has priority over vision.\n"
	"2. You must choose the most conservative interpretation.\n"
	"3. 'calm' is NOT a valid output.\n"
	"4. If unsure, prefer defensive or in pain.\n\n"
	"Allowed outputs ONLY:\n"
	"affectionate, angry, backing off, defensive, hungry, happy, "
	"hunting, in heat, calling kittens, in pain, wanting attention\n\n"
	"Answer format EXACTLY:\n"
	"The cat is [label]."
	)
	},
	{
	"role": "user",
	"content": (
	f"AUDIO SIGNAL (PRIMARY): {audio_ctx}\n"
	f"VISION OBSERVATIONS (SECONDARY): {vlm_desc}\n\n"
	"FINAL DECISION:"
	)
	}
	]

	input_text = llm_tok.apply_chat_template(
	messages,
	tokenize=False,
	add_generation_prompt=True
	)

	inputs = llm_tok(input_text, return_tensors="pt").to(DEVICE)

	with torch.no_grad():
	outputs = llm_model.generate(
	**inputs,
	max_new_tokens=15,
	do_sample=False,
	temperature=0.0,
	pad_token_id=llm_tok.eos_token_id,
	eos_token_id=llm_tok.eos_token_id
	)

	generated = llm_tok.decode(
	outputs[0][inputs["input_ids"].shape[1]:],
	skip_special_tokens=True
	).lower()

	for cat in CATEGORIES:
	if cat.replace("_", " ") in generated:
	return f"The cat is {cat.replace('_', ' ')}."

	# =====================================================
	# 5. FINAL FAILSAFE (NEVER CALM)
	# =====================================================
	return "The cat is defensive."


	# ==========================================
	# 3. PIPELINE ANALYSE COMPLETE (CORRIGÉ)
	# ==========================================
	@spaces.GPU(duration=120)
	def analyze_cat_v12_final(video_path):
	if not video_path:
	return "❌ Aucune vidéo.", None

	if torch.cuda.is_available():
	torch.cuda.empty_cache()

	tmp_audio = f"temp_{os.getpid()}_{int(time.time())}.wav"
	start_total = time.time()

	# --------------------------------------------------
	# Helper: clean VLM repetitions (cheap & mobile-safe)
	# --------------------------------------------------
	def clean_vlm_output(text):
	sentences = text.split(". ")
	cleaned = []
	seen = set()
	for s in sentences:
	key = s.strip().lower()
	if key and key not in seen:
	seen.add(key)
	cleaned.append(s.strip())
	return ". ".join(cleaned)

	try:
	# =========================
	# A. AUDIO
	# =========================
	t_0 = time.time()
	clip = VideoFileClip(video_path)
	audio_probs = np.zeros(len(CATEGORIES))

	if clip.audio:
	clip.audio.write_audiofile(tmp_audio, fps=16000, logger=None)
	w, _ = librosa.load(tmp_audio, sr=16000, duration=5.0)
	if len(w) < 48000:
	w = np.pad(w, (0, 48000 - len(w)))

	mel = librosa.feature.melspectrogram(y=w, sr=16000, n_mels=192)
	mel_db = (librosa.power_to_db(mel, ref=np.max) + 40) / 40
	img = cv2.resize(
	(np.vstack([mel_db, np.zeros((10, mel_db.shape[1]))]) * 255).astype(np.uint8),
	(224, 224)
	)

	img_t = (
	torch.tensor(img)
	.unsqueeze(0)
	.repeat(1, 3, 1, 1)
	.float()
	.to(DEVICE) / 255.0
	)

	with torch.no_grad():
	pa = F.softmax(audio_models['A'](img_t), dim=1)
	pb = F.softmax(audio_models['B'](img_t), dim=1)
	ic = audio_models['ast_ext'](
	w, sampling_rate=16000, return_tensors="pt"
	).to(DEVICE)
	pc = F.softmax(audio_models['C'](**ic).logits, dim=1)

	audio_probs = (
	pa * 0.3468 + pb * 0.2762 + pc * 0.3770
	).cpu().numpy()[0]

	clip.close()
	t_audio = time.time() - t_0

	# =========================
	# B. VISION (VLM STABILISÉ)
	# =========================
	t_1 = time.time()
	vlm_proc = AutoProcessor.from_pretrained(
	"HuggingFaceTB/SmolVLM2-256M-Video-Instruct"
	)

	vlm_prompt = (
	"You are a feline behavior expert.\n"
	"Describe ONLY observable physical features:\n"
	"- ears position\n"
	"- mouth state (open/closed/tense)\n"
	"- tail position or movement\n"
	"- body posture\n"
	"Use short factual sentences.\n"
	"One observation per sentence.\n"
	"Do NOT interpret mood."
	)

	messages = [
	{
	"role": "user",
	"content": [
	{"type": "video", "path": video_path},
	{"type": "text", "text": vlm_prompt}
	]
	}
	]

	vlm_inputs = vlm_proc.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt"
	).to(DEVICE)

	input_length = vlm_inputs["input_ids"].shape[1]

	with torch.no_grad():
	vlm_out = vlm_model.generate(
	**vlm_inputs,
	max_new_tokens=80,
	do_sample=False,
	temperature=0.0,
	repetition_penalty=1.15, # 🔑 anti-loop
	no_repeat_ngram_size=5, # 🔑 anti-phrase répétée
	pad_token_id=vlm_proc.tokenizer.eos_token_id,
	eos_token_id=vlm_proc.tokenizer.eos_token_id
	)

	gen_tokens = vlm_out[0][input_length:]
	vlm_clean = vlm_proc.batch_decode(
	[gen_tokens], skip_special_tokens=True
	)[0]

	vlm_clean = vlm_clean.strip().split("\n")[0]
	if vlm_clean.lower().startswith("assistant:"):
	vlm_clean = vlm_clean.split(":", 1)[-1].strip()

	# nettoyage final anti-répétition
	vlm_clean = clean_vlm_output(vlm_clean)

	if torch.cuda.is_available():
	torch.cuda.empty_cache()

	t_vlm = time.time() - t_1

	# =========================
	# C. JUGE
	# =========================
	t_2 = time.time()
	top_idx = np.argmax(audio_probs)
	audio_ctx = f"{CATEGORIES[top_idx].upper()} ({audio_probs[top_idx]*100:.1f}%)"
	judge_decision = call_peace_judge(audio_ctx, vlm_clean)
	t_llm = time.time() - t_2

	# =========================
	# D. VISUELS
	# =========================
	top5 = np.argsort(audio_probs)[-5:][::-1]
	fig = px.bar(
	x=[audio_probs[i] * 100 for i in top5],
	y=[CATEGORIES[i].upper() for i in top5],
	orientation="h",
	title="Top 5 Scores Audio",
	labels={"x": "Probabilité (%)", "y": "Émotion"},
	color=[audio_probs[i] * 100 for i in top5],
	color_continuous_scale="Viridis"
	)
	fig.update_layout(height=400, showlegend=False)

	# =========================
	# E. RAPPORT FINAL
	# =========================
	t_total = time.time() - start_total
	report = f"""⚖️ VERDICT JUGE : {judge_decision}
	------------------------------------------
	👁️ VISION : {vlm_clean}
	📊 AUDIO : {audio_ctx}
	⏱️ TEMPS : Audio {t_audio:.2f}s \| Vision {t_vlm:.2f}s \| Juge {t_llm:.2f}s \| Total {t_total:.2f}s"""

	return report, fig

	except Exception as e:
	return f"❌ Erreur : {str(e)}", None

	finally:
	if os.path.exists(tmp_audio):
	try:
	os.remove(tmp_audio)
	except:
	pass


	# --- Interface Gradio ---
	with gr.Blocks(theme=gr.themes.Soft()) as demo:
	gr.Markdown("# 🐱 CatSense v12.13 - Vision Pure Mode")
	gr.Markdown("✅ SmolVLM2-256M + SmolLM2-135M Juge + Audio Ensemble")

	with gr.Row():
	with gr.Column():
	video_input = gr.Video(label="Vidéo du chat")
	btn = gr.Button("🚀 ANALYSER", variant="primary", size="lg")
	with gr.Column():
	report_out = gr.Textbox(label="Résultat complet", lines=12, interactive=False)
	chart_out = gr.Plot(label="Distribution des émotions (Audio)")

	btn.click(analyze_cat_v12_final, inputs=video_input, outputs=[report_out, chart_out])

	demo.launch()